def tfslim_vgg16():
import tensorflow as tf
from nets import nets_factory
from preprocessing import vgg_preprocessing
from model_tools.activations.tensorflow import load_resize_image
tf.reset_default_graph()
image_size = 224
placeholder = tf.placeholder(dtype=tf.string, shape=[64])
preprocess_image = lambda image: vgg_preprocessing.preprocess_image(
image, image_size, image_size, resize_side_min=image_size)
preprocess = lambda image_path: preprocess_image(
load_resize_image(image_path, image_size))
preprocess = tf.map_fn(preprocess, placeholder, dtype=tf.float32)
model_ctr = nets_factory.get_network_fn('vgg_16',
num_classes=1001,
is_training=False)
logits, endpoints = model_ctr(preprocess)
session = tf.Session()
session.run(tf.initialize_all_variables())
return TensorflowSlimWrapper(identifier='tf-vgg16',
labels_offset=1,
endpoints=endpoints,
inputs=placeholder,
session=session)
def tfslim_custom():
from model_tools.activations.tensorflow import load_resize_image
import tensorflow as tf
slim = tf.contrib.slim
tf.compat.v1.reset_default_graph()
image_size = 224
placeholder = tf.compat.v1.placeholder(dtype=tf.string, shape=[64])
preprocess = lambda image_path: load_resize_image(image_path, image_size)
preprocess = tf.map_fn(preprocess, placeholder, dtype=tf.float32)
with tf.compat.v1.variable_scope('my_model', values=[preprocess]) as sc:
end_points_collection = sc.original_name_scope + '_end_points'
# Collect outputs for conv2d, fully_connected and max_pool2d.
with slim.arg_scope([slim.conv2d, slim.fully_connected, slim.max_pool2d],
outputs_collections=[end_points_collection]):
net = slim.conv2d(preprocess, 64, [11, 11], 4, padding='VALID', scope='conv1')
net = slim.max_pool2d(net, [5, 5], 5, scope='pool1')
net = slim.max_pool2d(net, [3, 3], 2, scope='pool2')
net = slim.flatten(net, scope='flatten')
net = slim.fully_connected(net, 1000, scope='logits')
endpoints = slim.utils.convert_collection_to_dict(end_points_collection)
session = tf.compat.v1.Session()
session.run(tf.compat.v1.initialize_all_variables())
return TensorflowSlimWrapper(identifier='tf-custom', labels_offset=0,
endpoints=endpoints, inputs=placeholder, session=session)
def init(identifier,
preprocessing_type,
image_size,
net_name=None,
labels_offset=1,
batch_size=64,
model_ctr_kwargs=None):
import tensorflow as tf
from nets import nets_factory
tf.compat.v1.reset_default_graph()
placeholder = tf.compat.v1.placeholder(dtype=tf.string,
shape=[batch_size])
preprocess = TFSlimModel._init_preprocessing(placeholder,
preprocessing_type,
image_size=image_size)
net_name = net_name or identifier
model_ctr = nets_factory.get_network_fn(net_name,
num_classes=labels_offset +
1000,
is_training=False)
logits, endpoints = model_ctr(preprocess, **(model_ctr_kwargs or {}))
if 'Logits' in endpoints: # unify capitalization
endpoints['logits'] = endpoints['Logits']
del endpoints['Logits']
session = tf.compat.v1.Session()
TFSlimModel._restore_imagenet_weights(identifier, session)
wrapper = TensorflowSlimWrapper(identifier=identifier,
endpoints=endpoints,
inputs=placeholder,
session=session,
batch_size=batch_size,
labels_offset=labels_offset)
wrapper.image_size = image_size
return wrapper